Automated scoring for Korean essays utilizing the neural pairwise contrastive regression along with Korean NLP models

Introduction Automated Essay Scoring (AES) research has increasingly adopted deep learning and transformer-based language models. However, most prior studies have focused on English writing and pointwise score prediction, leaving the applicability of ranking-aware AES frameworks to Korean essays underexplored. Methods This study examined the applicability of Neural Pairwise Contrastive Regression (NPCR) to Korean essay scoring and compared Korean-specific and multilingual pretrained language models within this framework. Using 4,418 Grade 11 essays from the AI-Hub Korean essay corpus, NPCR was implemented to jointly learn absolute scores and pairwise score differences. Essays were represented using transformer embeddings, including KLUE-RoBERTa-large and multilingual variants, and the models were trained with a modified pairing strategy to address negatively skewed score distributions. Model performance was evaluated using Quadratic Weighted Kappa (QWK) and Pearson correlation. Results The results showed strong agreement between NPCR-based predictions and human ratings, particularly when large Korean-specific pretrained language models were used. Performance also varied by prompt type, with higher accuracy observed for argumentative tasks. Discussion These findings support the applicability of NPCR to Korean essay scoring and highlight the importance of pretrained model choice in ranking-based AES. The results further suggest that Korean-specific pretrained language models may provide advantages for AES in Korean educational contexts.